Private AI

AI your firm can defend in an audit

An employee pastes a client's details into a free chatbot. Right now someone is doing it. The AI gives them a draft. Information covered by POPIA leaves your network and lands on a foreign server, under foreign law. You have no record of what was shared or who shared it.

Banning AI pushes staff to their phones, where you see even less. Your competitors get faster each quarter. Or you put AI inside your network. Faster answers than the free options. A record of what happened. When a regulator asks, you hand it over.

Capabilities

The Enclave: Private by Architecture

We deploy a complete AI capability inside your network, on hardware you control. Tailscale zero-trust networking keeps AI traffic off the public internet. We host models under data-residency and no-training guarantees. For the strictest mandates, the full stack runs entirely on your premises. Connects to your choice of cloud backbone: AWS Bedrock, GCP Vertex AI, Azure OpenAI, and the tooling your team uses.

The Airlock: PII Protection on Every Path

A redaction step on the chat box leaves your database queries unfiltered. Your uploaded documents too. The Airlock screens prompts, pasted text, file uploads, and database result sets. When our AI queries your systems, the filter processes each result set before the model reads a row. Sensitive values become consistent tokens. The model reasons about "{CLIENT_7}" without ever holding the real name. Real values restore after the model finishes, inside your network, on human approval. This filter runs as a small model on your hardware: millions of tokens screened for cents, no perceptible delay, each masking decision logged.

The Manager: Delegate Work, Not Configure

You talk to the Manager. In a sentence: "Pull last week's sales by region each Monday, compare to target, mail the summary to the exec team." The Manager drafts a plan you can inspect: which systems it touches, what it reads, what it sends. You approve it once. We save the plan to your playbook. It runs on schedule until a human approves a new version. Over months, you build a library of recurring jobs. The person who wrote the original request can leave or get promoted. The jobs execute the same way.

Audit by Default: Each Action Has a Receipt

Each AI task runs from a plan you approved. The system records documents read, drafts written, messages sent, who approved what, which data was involved, timestamps. Consequential actions require named-person approval. Those approvals appear in the record. A regulator asks "what did the AI do on March 3rd?" You produce the report. Evidence attached. We do not sell compliance in a box. Nobody can. We build controls with evidence behind them: the technical measures, the paper trail, the human oversight. Together they let you defend each decision.

Examples